In the swiftly advancing realm of artificial intelligence, generative AI tools showcase remarkable potential. Nevertheless, the rising awareness of their potential for harm is undeniable.
In VERSES, we’ve illustrated concerns related to generative AI tools using data from diverse sources. These concerns predominantly fall into three categories: quality control and data accuracy, ethical considerations, and technical challenges, often exhibiting a certain level of interconnection.
Credit: VERSES1) Bias In, …Bias Out
Theme: Quality Control & Accuracy
A significant challenge within generative AI revolves around its inclination to replicate biases inherent in the training data. Instead of mitigating biases, these tools frequently amplify or perpetuate them, casting doubts on the precision of their applications—potentially resulting in more profound ethical dilemmas.
2) The Black Box Problem
Theme: Ethical & Legal Considerations
Another significant hurdle in embracing generative AI is the lack of transparency in its decision-making processes. With thought processes that are often uninterpretable, these AI systems face challenges in explaining their decisions, especially when errors occur on critical matters.
It’s worth noting that this is a broader problem with all generative AI systems and not just generative tools.
3) Expensive to Build and to Maintain
Theme: Complexity & Technical Hurdles
The training of generative AI models, such as the large language model (LLM) ChatGPT, comes with an exorbitant price tag, often soaring into the millions of dollars due to the substantial computational power and infrastructure demands. As an illustration, former OpenAI CEO Sam Altman has verified that the training cost for ChatGPT-4 amounted to a staggering $100 million.
4) Mindless Parroting
Theme: Quality Control & Precision
While possessing advanced capabilities, generative AIs are confined by the data and patterns on which they were trained. This constraint manifests in outputs that might not fully embrace the depth of human knowledge or effectively address a spectrum of diverse scenarios.
AI models are also known to possess the inability to use jokes in other than a mindless and unfunny charade. (Grok, xAI’s chatbot on X, is changing that).
5) (Mis) Alignment with Human Values
Theme: Ethical & Legal Considerations
In contrast to humans, generative AIs lack the ability to contemplate the repercussions of their actions in accordance with human values.
Although instances like the innocuous AI-generated “Balenciaga Pope” may seem harmless, it’s crucial to acknowledge that deepfakes could be utilized for more malicious purposes, like disseminating false information during public health crises.
This underscores the necessity for additional frameworks to ensure the operation of these systems within ethical boundaries.
6) Power Hungry
Theme: Complexity & Technical Hurdles
The environmental implications of generative AI are impossible to ignore. Given that processing units consume significant power, models like ChatGPT can incur costs equivalent to powering 33,000 U.S. households, and a single inquiry can be 10 to 100 times more power-intensive than sending one email.
7) Hallucinations
Theme: Quality Control & Accuracy
Generative AI models have a tendency to generate false statements or images when confronted with data gaps, prompting concerns about the accuracy of their output and potential repercussions.
Illustratively, in a promotional video for Google Bard, the chatbot incorrectly claimed that the James Webb Space Telescope had captured the first images of a planet beyond Earth’s solar system.
8) Copyright & IP Infringement
Theme: Ethical & Legal Considerations
The ethical handling of data takes center stage, particularly as numerous generative AI tools unlawfully incorporate copyrighted work without obtaining consent, providing credit, or compensating artists and creators, thereby infringing on their rights.
OpenAI has recently implemented a compensation program named Copyright Shield. This initiative covers legal expenses related to copyright infringement suits for specific customer tiers, opting for this approach over the removal of copyrighted material from ChatGPT’s training dataset.
9) Static Information
Theme: Complexity & Technical Hurdles
Maintaining the currency of generative AI models demands significant computational resources and time, posing a formidable technical challenge. Nevertheless, certain models are structured to facilitate incremental updates, presenting a potential solution to this intricate issue.
